Research On The Identification Method Of Superfluous Material Based On Spectrogram And Transfer Learning

Sealed electronic components are widely used in the aerospace field or advanced weapons and equipment,but the loose particles inside the components pose a serious threat to the reliability of the system.It is of great value to carry out research on the material information of loose particles for improving the production process and preventing the generation of loose particles from the source.At present,the research on the identification of loose particles materials is mainly based on the signal characteristics of sound,ignoring the image characteristics of sound.In addition,due to the particularity of sealed electronic components,only a small amount of detection data of the components to be tested can be obtained,which cannot meet the requirements of machine learning model training.Aiming at the above problems,we proposed a loose particle material identification method based on spectrogram and transfer learning.We converted loose particle pulses into spectrograms,and made full use of spectrogram features to improve material identification performance.We made full use of the material datasets established by the research group over the years,learned knowledge from a large number of labeled samples that could be used for components under test,and transfered to new recognition tasks.Improvements were made at the model level to solve the problem of large differences in the distribution of source and target domains and low recognition accuracy in the process of transfer learning.In this thesis,spectrogram and transfer learning were used as the research core.In view of several problems encountered in the identification of loose particle materials,research was carried out in the aspects of material identification method,feature enhancement,data set imbalance,etc.The main research contents are divided into three parts:(1)A method for identifying the material of loose particles based on spectrogram was proposed.The sound signal detected by loose particles was converted into spectrogram by the multiple simultaneous compression transformation method,and the data imbalance was processed.The generator and discriminator networks were used to perform adversarial training until both reach a Nash equilibrium state.Generators were used to generate new data with the aim of addressing data imbalances.Extracting the time-frequency features of the spectrogram,designing the feature combination of the spectrogram feature and the time-frequency domain feature,and training and optimize the machine learning algorithm.The model made full use of the spectral features to improve its own recognition performance.(2)In the case of scarcity of loose particle material data,the deep neural network model was prone to overfitting,which leaded to the problem of reduced recognition accuracy and poor generalization ability.This study proposed a model-transfer-based method for identifying redundant objects.We designed a deep convolutional network for material recognition and added a group normalization layer.We transfered material recognition knowledge to new tasks by reusing deep neural networks trained on source domain datasets.In this way,the recognition ability of the model could be improved when the number of samples is scarce.(3)Aiming at the situation that the distribution of material signals in the source domain and target domain data sets was quite different,based on the model transfer,and improving from the model level,a material recognition method based on domain confrontation and multi-source heterogeneous information fusion was proposed.Using the strategy of weight transfer and adversarial training,the material recognition knowledge learned from the source domain was transferred to the target domain network,and the distribution difference between the two domains was reduced through adversarial training.A feature fusion network was designed to fuse the spectrogram signal features with the material sound signal features,in order to improve the material recognition accuracy.It could realize the feature migration of different domains and fully use of multisource heterogeneous information.